Side Network for Face Recognition Bounding Box Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face recognition systems do not optimize and align images effectively between face detection and recognition networks, leading to suboptimal performance due to different alignment approaches in each network, which affects the accuracy of face recognition.
Innovation Solution
A Joint Face Alignment and Recognition system is introduced, which includes a side network that adjusts the bounding box output from a face detection network, using a new loss function to refine the bounding box and improve alignment, thereby enhancing the accuracy of the face recognition network by aligning the input image for better feature extraction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If separate face detection and recognition networks are used with different alignment approaches, then each network can be optimized independently, but the overall face recognition accuracy deteriorates due to misalignment between networks
Solution Approach 1:
The patent combines face detection and face recognition networks into a unified system with a shared alignment module. The detection network and recognition network both receive input from the same alignment component, ensuring consistent bounding box adjustments and key point predictions across both networks, thereby resolving the misalignment issue while maintaining independent optimization capabilities through modular architecture
Solution Approach 2:
The alignment module serves multiple functions simultaneously: it performs bounding box adjustment for both detection and recognition networks, predicts facial key points, and optimizes input images for both downstream tasks. This multi-functional approach ensures consistent alignment across networks while improving overall system efficiency
2Device complexity
If standard bounding box output from face detection network is used directly, then the system complexity remains low, but the alignment precision deteriorates affecting feature extraction quality
Solution Approach 1:
The system performs preliminary alignment adjustments to the bounding box output from the detection network before passing it to the recognition network. The alignment module pre-processes the bounding box coordinates and key point predictions, ensuring that the recognition network receives properly aligned input data, thereby improving alignment precision without significantly increasing system complexity
Solution Approach 2:
The alignment module acts as an intermediary component between the detection network and recognition network. It receives the standard bounding box output from detection, processes it through alignment transformations based on predicted key points, and outputs the adjusted bounding box to recognition, thereby improving alignment precision while maintaining a clear modular system structure
Data Source
AI summary
By adding a side network to a face recognition network, output of early convolution blocks may be used to determine relative bounding box values. The relative bounding box values may be used to refine existing boundary box value with an eye on improving the generation, by the face recognition network, of embedding vectors.